LGFNet: Local-Global Fusion Network with Fidelity Gap Delta Learning for Multi-Source Aerodynamics

This paper proposes LGFNet, a novel deep learning framework that combines a spatial perception layer with self-attention mechanisms and a fidelity gap delta learning strategy to effectively fuse multi-source aerodynamic data, thereby achieving state-of-the-art accuracy in capturing both high-resolution local flow structures and global aerodynamic trends.

The Gist

Imagine you are trying to draw a perfect map of a mountainous landscape. You have two sources of information, but neither is perfect on its own:

1. The Satellite Map (CFD Data): This is a computer simulation. It covers the entire mountain range perfectly and shows you the general shape of the hills and valleys. However, because it's a simulation, it's a bit "blurry." It misses the sharp, jagged edges of the cliffs and the tiny details of the rocks.
2. The Hiker's Sketch (Wind Tunnel/Flight Test Data): This comes from real people measuring the mountain. It captures the exact sharpness of the cliffs and the tiny details perfectly. But, the hikers only measured a few specific spots. They didn't walk the whole mountain, so there are huge gaps in their sketch.

The Problem:
If you just use the Satellite Map, your map is smooth but inaccurate. If you try to fill in the gaps of the Hiker's Sketch using only the few points they took, you might draw wild, wiggly lines that don't make sense (like a cliff appearing where there is a valley).

The Solution: LGFNet
The paper introduces a new AI system called LGFNet (Local-Global Fusion Network) that acts like a master cartographer who can combine these two sources perfectly. It uses a clever three-step process to create a map that is both globally accurate (covers the whole mountain) and locally sharp (captures every jagged cliff).

Here is how it works, using simple analogies:

1. The "Sliding Window" (The Local Detective)

Imagine the AI puts a small magnifying glass over the map and slides it inch-by-inch across the terrain.

• What it does: This is the Spatial Perception Layer. It looks closely at small, local areas to find the "sharp edges" (like shock waves in air, which are like sudden cliffs in the data).
• Why it matters: Standard AI often smooths things out too much, turning a sharp cliff into a gentle hill. This sliding window ensures the AI doesn't miss the jagged, high-frequency details.

2. The "Self-Attention" (The Global Architect)

Now, imagine the AI steps back and looks at the whole mountain range at once, connecting the dots between the far-left valley and the far-right peak.

• What it does: This is the Relational Reasoning Layer. It uses a mechanism called "Self-Attention" to understand how one part of the map influences another part far away.
• Why it matters: Airflow is connected; what happens at the front of a plane affects the back. This layer ensures the map makes sense globally and doesn't have weird, isolated errors. It acts like a "noise filter," smoothing out the random mistakes in the hiker's sketch while keeping the big picture correct.

3. The "Delta Learning" (The Smart Correction)

This is the secret sauce. Instead of asking the AI to draw the entire mountain from scratch, the AI is told: "Don't redraw the mountain. Just fix the mistakes."

• The Strategy: The AI takes the blurry Satellite Map (the "Low-Fidelity" data) as a base. It then calculates the difference (the "Delta") between the blurry map and the few sharp points from the Hiker.
• The Result: The AI learns to predict only the errors and adds them to the base map. This ensures the final result keeps the smooth, logical flow of the computer simulation but adds the sharp, real-world accuracy of the measurements.

Why is this a big deal?

In the real world, this technology helps engineers design better airplanes.

• Before: They had to choose between a cheap, blurry computer model or an expensive, incomplete real-world test.
• Now: With LGFNet, they get the best of both worlds. They can predict exactly how air will flow over a new plane design with high precision, even in tricky situations like "shock waves" (sudden changes in air pressure), without needing to build a physical model for every single test.

In a nutshell:
LGFNet is like a smart editor who takes a rough draft (computer simulation) and a few handwritten notes (real tests) and produces a final book that is both factually perfect and beautifully detailed, fixing the errors without losing the original story.

Technical Summary

1. Problem Statement

In aerospace engineering, obtaining precise aerodynamic data is critical for design and flight control. However, existing data sources have distinct limitations:

• CFD (Computational Fluid Dynamics): Provides high-volume, wide-range data but suffers from model assumptions and computational limitations, leading to inaccuracies in capturing sharp discontinuities (e.g., shock waves).
• Wind Tunnel Tests: Offer controlled, relatively accurate data but are costly, time-consuming, and often sparse.
• Flight Tests: Provide the highest fidelity (ground truth) but are extremely expensive, risky, and limited in coverage.

The Core Challenge: Current Multi-Source Data Fusion (MDF) methods struggle to balance high-resolution local fidelity (capturing sharp shock waves and discontinuities) with wide-range global dependency (capturing long-range topological correlations). Traditional methods like Kriging fail to capture high-frequency non-linearities, while standard Deep Learning (DL) models often suffer from "texture bias" or act as low-pass filters that smooth out critical local features.

2. Methodology: LGFNet

The authors propose LGFNet (Local-Global Fusion Network), an end-to-end pyramid architecture designed for multi-scale feature decomposition. The framework integrates three core functional layers and a novel training strategy.

A. Fidelity Gap Delta Learning (FGDL) Strategy

Instead of predicting high-fidelity outputs directly, LGFNet treats low-fidelity CFD data as a "low-frequency carrier." The network is trained to predict the nonlinear discrepancy (residual) between the low-fidelity (CFD) and high-fidelity (Experimental) data.

• Formula: $y_H(x) = y_L(x) + \mathcal{R}_{gap}(x, y_L(x))$
• Benefit: This decouples geometry-governed base flows from high-frequency features, preventing unphysical smoothing while inheriting foundational physical trends from the simulation baseline.

B. Network Architecture

The network consists of three interdependent layers:

1. Spatial Perception Layer (SPL):

   • Mechanism: Utilizes a Sliding Window (SW) mechanism combined with a hierarchical CNN encoder.
   • Function: Reinforces the coverage and sequential continuity of local features. It preserves local gradient information (essential for detecting shock waves) by processing overlapping local blocks rather than standard batch processing.
   • Design: Uses anisotropic pooling to compress spatial dimensions while preserving aerodynamic feature dimensions, ensuring topological correlations (e.g., between Mach number and pressure coefficient) are maintained.

2. Relational Reasoning Layer (RRL):

   • Mechanism: Integrates Multi-Head Self-Attention (MHSA).
   • Function: Captures global topological structures and long-range dependencies (e.g., the correlation between leading-edge features and trailing-edge separation).
   • Role: Acts as a global low-pass filter to dynamically weigh data relationships and suppress interpolation noise introduced during data alignment, bridging distant but physically coupled aerodynamic states.

3. Feature Synthesis Layer (FSL):

   • Mechanism: A symmetric decoder with skip connections.
   • Function: Reconstructs high-resolution responses by fusing the globally reasoned features (from RRL) with the reinforced local spatial details (from SPL).
   • Benefit: Skip connections bridge the "semantic gap," ensuring that macro-scale topological structures do not compromise the resolution of sharp local features like shock waves.

C. Data Alignment

Before training, a Kriging interpolation model is used to align heterogeneous datasets (CFD and Experimental) onto a unified state space. The training range is restricted to the intersection of the numerical ranges of both datasets to avoid extrapolation errors.

3. Key Contributions

1. Specialized Framework: Proposes a three-layer architecture (SPL, RRL, FSL) specifically designed to simultaneously extract and integrate local discontinuities and global aerodynamic trends.
2. Hybrid Mechanism: Introduces a novel combination of Sliding Windows (for local continuity), Self-Attention (for global context), and Residual Learning (FGDL) to handle the "fidelity gap."
3. Fidelity Gap Delta Learning (FGDL): Reformulates the fusion task into a residual learning process, treating CFD as a baseline carrier to explicitly approximate nonlinear discrepancies, thereby preserving physical trends while correcting biases.
4. State-of-the-Art Performance: Demonstrates superior accuracy and uncertainty reduction across diverse aerodynamic scenarios compared to existing baselines.

4. Experimental Results

The model was validated on two distinct scenarios:

Scenario 1: RAE2822 Airfoil (Surface Pressure Distribution)

• Task: Fusing CFD pressure coefficients ($C_p$) with wind tunnel data across subsonic and transonic regimes.
• Key Findings:
  • Accuracy: LGFNet achieved the lowest RMSE (0.0591 in Case 1) and highest $R^2$ (0.9915) among all baselines (HK, DNN, MSFM, XGBoost, ArGEnT).
  • Shock Wave Capture: Unlike DNN (which smoothed shock waves) and ArGEnT (which over-smoothed them), LGFNet accurately reconstructed the position and intensity of shock wave discontinuities.
  • Uncertainty: LGFNet significantly reduced prediction uncertainty compared to raw experimental data, while avoiding the "overconfidence" (artificially low uncertainty with high error) seen in pure DNN models.

Scenario 2: CARDC Aircraft (High-Dimensional Force Coefficients)

• Task: Fusing CFD data with sparse flight test data for axial ($C_x$), normal ($C_y$), and lateral ($C_z$) force coefficients.
• Key Findings:
  • Robustness: LGFNet achieved the lowest RMSE for the challenging lateral force coefficient ($C_z$: 0.0169), outperforming all baselines.
  • Noise Suppression: The model effectively suppressed high-frequency noise in flight test data while adhering to the smooth physical trends of CFD simulations.
  • Ablation Studies: Confirmed that removing either the Sliding Window (local) or Attention (global) modules significantly degraded performance, proving the necessity of the dual-architecture.

3D Visualization

Visualizations of the fused surfaces demonstrated that LGFNet successfully corrected the systematic bias of CFD data (aligning vertically with flight test data) while preserving the smooth geometric topology and trends of the CFD baseline.

5. Significance

• Engineering Impact: LGFNet provides a robust paradigm for integrating low-cost, high-volume CFD data with sparse, high-fidelity experimental data. This enables the creation of comprehensive, high-accuracy aerodynamic databases without the prohibitive cost of full-scale flight testing.
• Scientific Advancement: The paper addresses a critical gap in deep learning for fluid dynamics: the inability of standard attention mechanisms to preserve sharp local discontinuities. By combining local inductive biases (sliding windows) with global reasoning (attention), it offers a new solution for multi-scale physical modeling.
• Uncertainty Quantification: The method not only improves point predictions but also provides reliable uncertainty estimates, crucial for risk assessment in aerospace design.

In conclusion, LGFNet represents a significant step forward in data-driven aerodynamics, offering a balanced solution that captures both the "sharpness" of local flow structures and the "breadth" of global aerodynamic trends.